Learning and inferring the diurnal variability of cyanobacterial blooms from high-frequency time-series satellite-based observations.

ABSTRACT

Observational evidences have suggested that the surface scums of cyanobacterial harmful blooms (CyanoHABs) are highly patchy, and their spatial patterns can vary significantly within hours. This stresses the need for the capacity to monitor and predict their occurrence with better spatiotemporal continuity, in order to understand and mitigate their causes and impacts. Although polar-orbiting satellites have long been used to monitor CyanoHABs, these sensors cannot be used to capture the diurnal variability of the bloom patchiness due to their long revisit periods. In this study, we use the Himawari-8 geostationary satellite to generate high-frequency time-series observations of CyanoHABs on a sub-daily basis not possible from previous satellites. On top of that, we introduce a spatiotemporal deep learning method (ConvLSTM) to predict the dynamics of bloom patchiness at a lead time of 10 min. Our results show that the bloom scums were highly patchy and dynamic, and the diurnal variability was assumed to be largely associated with the migratory behavior of cyanobacteria. We also show that, ConvLSTM displayed fairly satisfactory performance with promising predictive capability, with Root Mean Square Error (RMSE) and determination coefficient (R2) varying between 0.66∼1.84 µg/L and 0.71∼0.94, respectively. This suggests that, by adequately capturing spatiotemporal features, the diurnal variability of CyanoHABs can be well learned and inferred by ConvLSTM. These results may have important practical implications, because they suggest that spatiotemporal deep learning integrated with high-frequency satellite observations could provide a new methodological paradigm in nowcasting of CyanoHABs.